As of Q1 2020, CBRE EA is updating its forecasting methodology by integrating our market and submarket forecasts into a single multilevel model.
Our previous methodology dealt with the hierarchical nature of our data by using separate models for the market and submarket levels. It accounted for the multilevel structure of the data by using two models and a top-down approach to aggregate consistency.
First, we aggregated all data to the top level: the market. Market fundamentals were forecasted using a vector autoregression (VAR), estimating the interdependencies among rent, vacancy, and stock, and projecting those variables into the future. Second, the just-forecasted market numbers were broken into portions and allocated to each submarket using a structural model which used short- and long-run assumptions to ensure stability and aggregate consistency.
The top-down model works well in most situations. The assumptions of the second stage restrict it from quickly changing the supply or demand allocated to a submarket. This insulates the forecast against local noise, but, when a submarket sees a construction boom, this constraint can fail to allocate enough demand, leading to a vacancy spike.
Submarkets are difficult to forecast. They have fewer data points than markets. This makes them more erratic, noisier, and sensitive to outliers. A new addition to a submarket can have a large impact on predictions.
The new method uses similar equations as the former method but combines them into one model. This combined model gives us more flexibility in forecasting submarkets — particularly submarkets which are rapidly adding new stock. The result is a new model that is more responsive at both the market and submarket levels, pooling information between the two levels while remaining consistent with long-run market dynamics.
Our new method allows us to account for flexibility and stability. Using a multilevel modelling framework, we estimate both the submarket and market levels simultaneously with a panel vector autoregression. Submarkets are included in the model as a distribution of deviations from the market. Submarkets are freely estimated, but they are tied to the market.
This formulation mirrors reality: submarkets depend on both one another and the overall market. By pooling information from all levels in one step, the stability of the market-level variables, such as employment, serve as a scaffold to guide noisy submarket forecasts. This provides enough structure to keep the forecasts stable in the long-run, while allowing individual submarkets to respond quickly to changing neighborhood characteristics in the short-run.